Air conditioning user comfort evaluation method considering meteorological influence

Abstract

The present invention discloses a method for evaluating air conditioning user comfort that takes into account meteorological influences. The method comprises step S1: collecting comfort data of air conditioning users under different meteorological conditions; constructing and annotating a dataset based on the data; step S2: constructing a scoring model, which includes a feature extraction module and a prediction module based on a collaborative firefly algorithm. The data from the dataset in step S1 is imported into the feature extraction module to obtain preliminary features; step S3: importing the preliminary features into the prediction module to obtain a predicted comfort evaluation result; and step S4: dynamically adjusting air conditioning control parameters based on real-time feedback of the comfort evaluation result. The present invention constructs a complete comfort evaluation system based on air conditioning setting parameters, meteorological data, and user physiological response data. It uses causal inference modeling to clarify the path of weather-related impact on comfort and improve data reliability.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence model training methods, and specifically relates to an air-conditioning user comfort evaluation method taking meteorological influences into consideration. Background Art

[0002] With the rapid development of smart home technology, intelligent control of air-conditioning systems has gradually become a research hotspot. However, existing air-conditioning control strategies are mainly based on fixed rules or simple machine learning methods, which make it difficult to accurately adapt to the comfort needs of different users under different environmental conditions. Especially under complex meteorological conditions, traditional methods often ignore the combined impact of environmental factors on air-conditioning settings and human physiological state, resulting in poor accuracy and adaptability of comfort prediction. In addition, while most current deep learning-based comfort assessment methods can extract high-dimensional features from data, they often lack causal relationship modeling and may suffer from spurious correlations, resulting in insufficient model generalization ability. Furthermore, deep learning models are prone to falling into local optimality during training, resulting in unstable training and affecting the reliability of comfort prediction.

[0003] The Chinese invention patent with publication number CN119509013A proposes a comfort evaluation method, device, equipment and storage medium for an air conditioner. The present invention obtains the current operating data of the air conditioner and user operation behavior information, and performs comfort evaluation on the current operating data and user operation behavior information through a trained comfort evaluation model to obtain comfort evaluation data. Finally, the current comfort evaluation result of the air conditioner is determined based on a pre-set comfort measurement interval combined with the comfort evaluation data, thereby avoiding the technical problem of low accuracy of comfort evaluation when users use air conditioners in the existing technology and improving the comprehensiveness of comfort evaluation.

[0004] A Chinese invention patent, publication number CN119477038A, proposes a game-theory-based energy-saving evaluation method for wind-heat storage systems. The method includes the following steps: establishing an evaluation index system model, including three indicators: energy efficiency, energy structure, and pollutant emissions per unit of electricity; establishing a GTCE-TOPSIS evaluation method model, using the subjective weighting method (AHP) and the objective weighting method (EWM) to determine the weights of the evaluation index values, and aggregating the weighted results; constructing a weighted matrix, calculating the positive and negative ideal solutions, as well as the distances between the positive and negative ideal solutions, for each object to be evaluated in the weighted matrix; then calculating the proximity values ​​of the objects to be evaluated; and finally, grading the proximity values ​​to obtain a final grade. While providing users with a comfortable indoor environment, this method conserves energy and promotes the consumption of clean energy, thereby further achieving the clean development goals of the energy system.

[0005] The Chinese invention patent with publication number CN119469648A proposes a cruise ship theater temperature and airflow organization evaluation system, including: a regulation module for adjusting the air supply speed, air supply volume and air supply temperature in the cruise ship theater; a data acquisition module for collecting the air supply speed, air supply volume and air supply temperature data of the passenger sightseeing area and the performance area; a data recording module for recording the collected air supply speed, air supply volume and air supply temperature data in real time to obtain the temperature and airflow change process over time; a data evaluation module for obtaining the indoor temperature field and airflow organization distribution in a stable state based on the obtained temperature and airflow change process over time. A cruise ship theater temperature and airflow organization evaluation method is also disclosed. The present invention improves the measurement accuracy of the temperature and airflow field of cruise ship theaters, provides data support for the subsequent adjustment of the air conditioning vent position, and improves the theater's environmental comfort and passenger satisfaction.

[0006] The existing technology has the following deficiencies:

[0007] Only neural networks are used for comfort prediction without considering the causal relationship of data, resulting in pseudo-correlation problems; neural networks are prone to falling into local optimality, unstable training, and gradient vanishing or gradient exploding problems; existing methods mostly rely on black-box deep learning models and lack interpretability; there is a lack of optimized feedback mechanism for intelligent air conditioning control; feature extraction is single and does not consider the fusion of multi-source data; the neural network structure is fixed and difficult to adapt to different environments and data changes. Summary of the Invention

[0008] In response to the technical problems in the existing technology, a method for evaluating air-conditioning user comfort considering meteorological influences is provided, including:

[0009] Step S1: Collect the comfort data of air-conditioning users under different meteorological conditions; construct and annotate the data set;

[0010] Step S2: Construct a scoring model, which includes a feature extraction module and a prediction module based on the collaborative firefly algorithm. The data from the dataset in step S1 is imported into the feature extraction module to obtain preliminary features. During this process, a causal inference fusion gate is used, and a differentiable causal graph network is used to model the causal chain of comfort levels corresponding to different source collection information, suppressing the interference of spurious correlation features.

[0011] Step S3: Import the preliminary features into the prediction module to obtain the predicted comfort evaluation results;

[0012] Step S4: Dynamically adjust the air conditioning control parameters based on the real-time feedback comfort evaluation results.

[0013] Furthermore, step S2 is specifically as follows:

[0014] Step S21: Initialize the parameters of the neural network of the feature extraction module. The initialization method is expressed as follows:

[0015] ;

[0016] ;

[0017] Where, is the initial weight matrix of the neural network of the feature extraction module, is a random initialization function, is the initial bias vector of the neural network of the feature extraction module, and are the dimensions of the input and output layers of the neural network of the feature extraction module respectively;

[0018] Step S22: Optimize the parameters of the neural network of the feature extraction module using the collaborative firefly algorithm. Each firefly represents a set of network parameters, including the weight matrix and bias vector of the neural network of the feature extraction module. The loss function is used as the brightness of the fitness to move and adjust in the search space. Specifically, the collaborative firefly algorithm optimizes the neural network of the feature extraction module as follows:

[0019] ;

[0020] in, For the The brightness of a firefly, is the loss function of the neural network of the feature extraction module, and are the input comfort evaluation data and labels of the neural network of the feature extraction module corresponding to the i-th group of network parameters, is the weight matrix of the neural network of the feature extraction module, is the bias vector of the neural network of the feature extraction module;

[0021] The position of the firefly is updated based on its brightness. The update method is expressed as follows:

[0022] ;

[0023] Where, is the updated position of the firefly, corresponding to the updated weight of the neural network of the feature extraction module; is the attraction coefficient, is the light absorption coefficient; For comparison The positions of other brighter fireflies are characterized by Compare The corresponding brightness function calculation value is higher; For the The location of the fireflies, For the The location of the fireflies, ;

[0024] Step S23: The input comfort evaluation data of the neural network corresponding to the group network parameters The system collects information from different sources, including meteorological data, air conditioning settings, and physiological response data. It uses a causal inference fusion gate and a differentiable causal graph network to model the causal chain of meteorological data → air conditioning settings → physiological response data → comfort, suppressing interference from spurious correlation features and thus obtaining preliminary features.

[0025] Step S24: Through neural network architecture search, different neural network architectures as feature extraction modules are automatically tested. The network architecture includes the number of layers, number of neurons, and activation function type to determine the optimal structure. The neural network architecture search determines the optimal structure through cross-validation, which is expressed as:

[0026] ;

[0027] Where, Search function for network structure, is the comfort evaluation dataset after segmentation, is the number of cross-validation folds, Represents the k-fold cross validation under a given weight matrix Wp and the k-fold comfort evaluation dataset The performance evaluation value under The loss of fold cross validation, represents the minimum evaluation function;

[0028] Step S25: Evaluate the optimized neural network used as the feature extraction module, specifically by comparing the correlation between the extracted preliminary features and the labels. The evaluation method is expressed as follows:

[0029] ;

[0030] Where, is a feature quality indicator, is the neural network feature extraction function, is the similarity calculation function, Represents preliminary features, Indicates a preset label;

[0031] Step S26: According to the feature quality index, the parameters of the firefly algorithm are adjusted to more finely adjust the parameters of the neural network. The adjustment method is expressed as follows:

[0032] ;

[0033] Where, is the adjusted attractiveness coefficient; is the attractive learning rate, represents the exponential function;

[0034] Step S27: Repeat steps S21 to S26 until a preset stopping condition is met.

[0035] Furthermore, the loss function in step S22 Expressed as:

[0036] ;

[0037] Where, The number of neural networks that feed the feature extraction module for the current batch, Indicates the category with the highest probability obtained by calculating the preset Softmax classification function based on the neural network prediction output of the feature extraction module; represents the true label of the comfort data; is the counterfactual regularization term; is the regularization parameter of the neural network of the feature extraction module; is the weight matrix The Frobenius norm of .

[0038] Furthermore, in step S24, the optimal structure is determined by cross-validation. Fold cross validation, expressed as:

[0039] ;

[0040] Where, Represents the k-fold cross validation under a given weight matrix Wp and the k-fold comfort evaluation dataset The performance evaluation value under Indicates the The number of comfort evaluation data points in the comfort evaluation dataset; Represents the loss function for neural network architecture search; Is the neural network in a given weight matrix Lower pair input The output of is calculated by the preset Softmax classification function to obtain the category with the highest probability. Represents the output of the neural network.

[0041] Furthermore, step S23 specifically includes: constructing a differentiable causal graph network, defining structured causal equations between multi-source data of source collected information, and explicitly modeling causal paths through a gating mechanism; specifically:

[0042] Get and preset weather data , the air conditioning setting parameters are , the physiological response data is ;

[0043] in, is the meteorological data dimension; Set the dimensions for the air conditioner, Dimensions representing physiological response data; represents the set of real numbers;

[0044] The causal propagation process of the differentiable causal graph network is modeled as:

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] Where, The weight matrix of the causal path from meteorological data to air conditioning, The bias term for the causal path from meteorological data to air conditioning, The weight matrix for the causal paths from conditioning to physiological response data, A bias term that sets the causal path to the physiological response data for conditioning, is the gating weight matrix, is the gate dimension, is the gate bias term, is the ReLU activation function, is the Sigmoid function, is the vector concatenation operation, is element-wise multiplication, is a dynamic gating vector used to control the weight of the indirect impact of meteorological data on physiological response data; represents the causal inference fusion gate;

[0050] The pseudo-correlation suppression mechanism constrains the feature space through causal intervention loss, eliminates the influence of confounding factors, and defines the counterfactual regularization term as:

[0051] ;

[0052] Where, is the comfort prediction value of the neural network, For meteorological data The Jacobian matrix of the preliminary features, Set up the air conditioner The Jacobian matrix of the preliminary features, is the Frobenius norm, the square root of the sum of the squares of the matrix elements, is the regularization penalty coefficient, represents the set of positive real numbers;

[0053] By constraining the direct gradient contribution of meteorological data and air conditioning settings to the initial features of the output, the differentiable causal graph network is made to be only based on physiological response data. Predict the causal path to block the non-causal association path;

[0054] Multi-source feature fusion is performed by performing a residual connection between the output of the differentiable causal graph network and the original features, preserving the complementary information of explicit causal features and implicit correlation features. Through the dynamic fusion coefficient, the contribution of explicit causal reasoning and data-driven implicit features is balanced to avoid oversimplification of real scenarios by causal modeling. It is expressed as:

[0055] ;

[0056] Where, is the original input feature, It is a multi-layer perceptron used to extract implicit features; is the learnable dynamic fusion coefficient, and the update method is , is the learning rate, is the feature quality indicator, is the final output of the differentiable causal graph network, that is, the preliminary feature.

[0057] Furthermore, step S3 is specifically as follows:

[0058] The preliminary features are input into the prediction module, which is a fully connected layer. The prediction probability of each comfort level is output through the Softmax classification function. The category with the highest probability is used as the final comfort evaluation result to reflect the user's comfort status in the current air-conditioning environment.

[0059] The positive progress effect of the present invention is:

[0060] 1. The present invention constructs a complete comfort evaluation system based on air conditioning setting parameters, meteorological data and user physiological response data, adopts causal inference modeling, clarifies the impact path of meteorological conditions on comfort, and improves data reliability.

[0061] 2. The present invention optimizes the neural network based on the collaborative firefly algorithm. It optimizes the weights and biases of the neural network through the global search capability of the firefly algorithm, overcomes the problems of gradient disappearance, gradient explosion and local optimality, and adopts neural network architecture search to automatically optimize the neural network structure and improve the generalization ability of the model.

[0062] 3. This paper adopts a causal inference fusion mechanism to model the causal chain of "meteorological data → air conditioning settings → physiological response data → comfort level", thereby eliminating the influence of spurious correlations and improving the interpretability and reliability of the model. At the same time, through the causal intervention loss constraint (counterfactual regularization term), it avoids the overfitting problem caused by data-driven methods.

[0063] 4. The present invention combines causal features and data-driven implicit features, adopts residual connection for feature fusion, improves prediction accuracy, and dynamically adjusts air-conditioning parameters according to comfort prediction results through intelligent feedback mechanism to enhance user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The present invention is a flowchart of the steps of a method for evaluating air-conditioning user comfort taking into account meteorological influences.

[0065] Figure 2 This is a comparison chart of the accuracy of the air-conditioning user comfort evaluation method considering meteorological influences of the present invention and different algorithms in the prior art under training samples of different sizes.

[0066] Figure 3 This is a comparison chart of the characteristic quality of an air-conditioning user comfort evaluation method considering meteorological influences of the present invention and different algorithms in the prior art at different noise levels based on the average trend line. DETAILED DESCRIPTION

[0067] The following describes the embodiments of the present invention through specific examples. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention.

[0068] A method for evaluating air-conditioning user comfort considering meteorological influences includes:

[0069] Step S1: Collect the comfort data of air-conditioning users under different meteorological conditions; construct and annotate the data set;

[0070] Step S2: Construct a scoring model, which includes a feature extraction module and a prediction module based on the collaborative firefly algorithm. The data from the dataset in step S1 is imported into the feature extraction module to obtain preliminary features. During this process, a causal inference fusion gate is used, and a differentiable causal graph network is used to model the causal chain of comfort levels corresponding to different source collection information, suppressing the interference of spurious correlation features.

[0071] Step S3: Import the preliminary features into the prediction module to obtain the predicted comfort evaluation results;

[0072] Step S4: Dynamically adjust the air conditioning control parameters based on the real-time feedback comfort evaluation results.

[0073] Furthermore, step S2 is specifically as follows:

[0074] Step S21: Initialize the parameters of the neural network of the feature extraction module. The initialization method is expressed as follows:

[0075] ;

[0076] ;

[0077] Where, is the initial weight matrix of the neural network of the feature extraction module, is a random initialization function, is the initial bias vector of the neural network of the feature extraction module, and are the dimensions of the input and output layers of the neural network of the feature extraction module respectively;

[0078] Step S22: Optimize the parameters of the neural network of the feature extraction module using the collaborative firefly algorithm. Each firefly represents a set of network parameters, including the weight matrix and bias vector of the neural network of the feature extraction module. The loss function is used as the brightness of the fitness to move and adjust in the search space. Specifically, the collaborative firefly algorithm optimizes the neural network of the feature extraction module as follows:

[0079] ;

[0080] in, For the The brightness of a firefly, is the loss function of the neural network of the feature extraction module, and are the input comfort evaluation data and labels of the neural network of the feature extraction module corresponding to the i-th group of network parameters, is the weight matrix of the neural network of the feature extraction module, is the bias vector of the neural network of the feature extraction module;

[0081] The position of the firefly is updated based on its brightness. The update method is expressed as follows:

[0082] ;

[0083] Where, is the updated position of the firefly, corresponding to the updated weight of the neural network of the feature extraction module; is the attraction coefficient, is the light absorption coefficient; For comparison The positions of other brighter fireflies are characterized by Compare The corresponding brightness function calculation value is higher; For the The location of the fireflies, For the The location of the fireflies, ;

[0084] Step S23: The input comfort evaluation data of the neural network corresponding to the group network parameters The system collects information from different sources, including meteorological data, air conditioning settings, and physiological response data. It uses a causal inference fusion gate and a differentiable causal graph network to model the causal chain of meteorological data → air conditioning settings → physiological response data → comfort, suppressing interference from spurious correlation features and thus obtaining preliminary features.

[0085] Step S24: Through neural network architecture search, different neural network architectures as feature extraction modules are automatically tested. The network architecture includes the number of layers, number of neurons, and activation function type to determine the optimal structure. The neural network architecture search determines the optimal structure through cross-validation, which is expressed as:

[0086] ;

[0087] Where, Search function for network structure, is the comfort evaluation dataset after segmentation, is the number of cross-validation folds, Represents the k-fold cross validation under a given weight matrix Wp and the k-fold comfort evaluation dataset The performance evaluation value under The loss of fold cross validation, represents the minimum evaluation function;

[0088] Step S25: Evaluate the optimized neural network used as the feature extraction module, specifically by comparing the correlation between the extracted preliminary features and the labels. The evaluation method is expressed as follows:

[0089] ;

[0090] Where, is a feature quality indicator, is the neural network feature extraction function, is the similarity calculation function, Represents preliminary features, Indicates a preset label;

[0091] Step S26: According to the feature quality index, the parameters of the firefly algorithm are adjusted to more finely adjust the parameters of the neural network. The adjustment method is expressed as follows:

[0092] ;

[0093] Where, is the adjusted attractiveness coefficient; is the attractive learning rate, represents the exponential function;

[0094] Step S27: Repeat steps S21 to S26 until a preset stopping condition is met.

[0095] Furthermore, the loss function in step S22 Expressed as:

[0096] ;

[0097] Where, The number of neural networks that feed the feature extraction module for the current batch, Indicates the category with the highest probability obtained by calculating the preset Softmax classification function based on the neural network prediction output of the feature extraction module; represents the true label of the comfort data; is the counterfactual regularization term; is the regularization parameter of the neural network of the feature extraction module; is the weight matrix The Frobenius norm of .

[0098] Furthermore, in step S24, the optimal structure is determined by cross-validation. Fold cross validation, expressed as:

[0099] ;

[0100] Where, Represents the k-fold cross validation under a given weight matrix Wp and the k-fold comfort evaluation dataset The performance evaluation value under Indicates the The number of comfort evaluation data points in the comfort evaluation dataset; Represents the loss function for neural network architecture search; Is the neural network in a given weight matrix Lower pair input The output of is calculated by the preset Softmax classification function to obtain the category with the highest probability. Represents the output of the neural network.

[0101] Furthermore, step S23 specifically includes: constructing a differentiable causal graph network, defining structured causal equations between multi-source data of source collected information, and explicitly modeling causal paths through a gating mechanism; specifically:

[0102] Get and preset weather data , the air conditioning setting parameters are , the physiological response data is ;

[0103] in, is the meteorological data dimension; Set the dimensions for the air conditioner, Dimensions representing physiological response data; represents the set of real numbers;

[0104] The causal propagation process of the differentiable causal graph network is modeled as:

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] Where, The weight matrix of the causal path from meteorological data to air conditioning, The bias term for the causal path from meteorological data to air conditioning, The weight matrix for the causal paths from conditioning to physiological response data, A bias term that sets the causal path to the physiological response data for conditioning, is the gating weight matrix, is the gate dimension, is the gate bias term, is the ReLU activation function, is the Sigmoid function, is the vector concatenation operation, is element-wise multiplication, is a dynamic gating vector used to control the weight of the indirect impact of meteorological data on physiological response data; represents the causal inference fusion gate;

[0110] The pseudo-correlation suppression mechanism constrains the feature space through causal intervention loss, eliminates the influence of confounding factors, and defines the counterfactual regularization term as:

[0111] ;

[0112] Where, is the comfort prediction value of the neural network, For meteorological data The Jacobian matrix of the preliminary features, Set up the air conditioner The Jacobian matrix of the preliminary features, is the Frobenius norm, the square root of the sum of the squares of the matrix elements, is the regularization penalty coefficient, represents the set of positive real numbers;

[0113] By constraining the direct gradient contribution of meteorological data and air conditioning settings to the initial features of the output, the differentiable causal graph network is made to be only based on physiological response data. Predict the causal path to block the non-causal association path;

[0114] Multi-source feature fusion is performed by performing a residual connection between the output of the differentiable causal graph network and the original features, preserving the complementary information of explicit causal features and implicit correlation features. Through the dynamic fusion coefficient, the contribution of explicit causal reasoning and data-driven implicit features is balanced to avoid oversimplification of real scenarios by causal modeling. It is expressed as:

[0115] ;

[0116] Where, is the original input feature, It is a multi-layer perceptron used to extract implicit features; is the learnable dynamic fusion coefficient, and the update method is , is the learning rate, is the feature quality indicator, is the final output of the differentiable causal graph network, that is, the preliminary feature.

[0117] Furthermore, step S3 is specifically as follows:

[0118] The preliminary features are input into the prediction module, which is a fully connected layer. The prediction probability of each comfort level is output through the Softmax classification function. The category with the highest probability is used as the final comfort evaluation result to reflect the user's comfort status in the current air-conditioning environment.

[0119] In order to verify the effectiveness of the technology of the present invention, the following experiments were conducted:

[0120] Reference Figure 2 In order to verify the adaptability and prediction accuracy of the proposed method under different data scales, the algorithm is compared with conventional technologies such as traditional neural networks, random forests and support vector machines to reveal its advantages in complex multi-source data fusion. The experiment uses the amount of training data as a variable to compare the accuracy trends of various methods in user comfort assessment. The results show that the proposed method exhibits higher learning efficiency and stable performance under different data scales, especially under limited data conditions, it can still maintain high prediction accuracy. This shows that the parameter search mechanism optimized by the collaborative firefly algorithm and the design of the causal inference fusion gate, compared with the simple reliance of traditional methods on explicit features, can effectively capture the deep correlation between meteorological data, air conditioning settings and physiological response data through dynamic balance causal path modeling and implicit feature extraction, thereby showing more significant performance improvement when the data scale is expanded, verifying the superiority of the algorithm in feature expression and pattern mining.

[0121] Reference Figure 3 To evaluate the feature extraction quality of different methods under noise interference, the robustness of key features such as meteorological parameters (i.e., meteorological data), device settings (i.e., air conditioning settings), and physiological signals (i.e., physiological response data) was compared. The experiments simulated sensor acquisition errors and user feedback noise to compare the methods' ability to preserve causal relationships between features. The results demonstrated that this technique can stably extract highly discriminative features despite noise interference. The differentiable causal graph network explicitly models the causal chain from "meteorological data-air conditioning settings-physiological response data," while counterfactual regularization suppresses spurious correlations. Compared to traditional neural networks' reliance on superficial statistical correlations, this method strengthens the propagation weights of causal paths through a dynamic gating mechanism, significantly outperforming the comparative methods in quality metrics for core features such as physiological signals. Even under extreme noise conditions, the model retains valid information from the original data through residual connections, demonstrating the algorithm's practical value in complex real-world scenarios.

[0122] The present invention has been described in detail above with reference to the embodiments of the accompanying drawings. A person skilled in the art can make various modifications to the present invention based on the above description. Therefore, certain details in the embodiments should not be construed as limiting the present invention. The scope of protection of the present invention shall be determined by the scope defined in the appended claims.

Claims

1. A method for evaluating air-conditioning user comfort considering meteorological influences, characterized in that: include: Step S1: collecting comfort data of air-conditioning users under different meteorological conditions; And use this to build a data set and annotate it; Step S2: Construct a scoring model, which includes a feature extraction module and a prediction module based on the collaborative firefly algorithm. The data from the dataset in step S1 is imported into the feature extraction module to obtain preliminary features. During this process, a causal inference fusion gate is used, and a differentiable causal graph network is used to model the causal chain of comfort levels corresponding to different source collection information, suppressing the interference of spurious correlation features. Step S3: Import the preliminary features into the prediction module to obtain the predicted comfort evaluation results; Step S4: Dynamically adjusting the air conditioning control parameters based on the real-time feedback comfort evaluation results; Step S2 is specifically as follows: Step S21: Initialize the parameters of the neural network of the feature extraction module. The initialization method is expressed as follows: ； ； Where, is the initial weight matrix of the neural network of the feature extraction module, is the random initialization function, is the initial bias vector of the neural network of the feature extraction module, and are the dimensions of the input and output layers of the neural network of the feature extraction module respectively; Step S22: Optimize the parameters of the neural network of the feature extraction module using the collaborative firefly algorithm. Each firefly represents a set of network parameters, including the weight matrix and bias vector of the neural network of the feature extraction module. The loss function is used as the brightness of the fitness to move and adjust in the search space. Specifically, the collaborative firefly algorithm optimizes the neural network of the feature extraction module as follows: ； in, For the The brightness of a firefly, is the loss function of the neural network of the feature extraction module, and are the input comfort evaluation data and labels of the neural network of the feature extraction module corresponding to the i-th group of network parameters, is the weight matrix of the neural network of the feature extraction module, is the bias vector of the neural network of the feature extraction module; The position of the firefly is updated based on its brightness. The update method is expressed as follows: ； Where, is the updated position of the firefly, corresponding to the updated weight of the neural network of the feature extraction module; is the attraction coefficient, is the light absorption coefficient; For comparison The positions of other brighter fireflies are represented by Compare The corresponding brightness function calculation value is higher; For the The location of the fireflies, For the The location of the fireflies, ; Step S23: The input comfort evaluation data of the neural network corresponding to the group network parameters The system collects information from different sources, including meteorological data, air conditioning settings, and physiological response data. It uses a causal inference fusion gate and a differentiable causal graph network to model the causal chain of meteorological data → air conditioning settings → physiological response data → comfort, suppressing interference from spurious correlation features and thus obtaining preliminary features. Step S24: Through neural network architecture search, different neural network architectures as feature extraction modules are automatically tested. The network architecture includes the number of layers, number of neurons, and activation function type to determine the optimal structure. The neural network architecture search determines the optimal structure through cross-validation, which is expressed as: ； Where, Search function for network structure, is the comfort evaluation dataset after segmentation, is the number of cross-validation folds, Represents the k-fold cross validation under a given weight matrix Wp and the k-fold comfort evaluation dataset The performance evaluation value under The loss of fold cross validation, represents the minimum evaluation function; Step S25: Evaluate the optimized neural network used as the feature extraction module, specifically by comparing the correlation between the extracted preliminary features and the labels. The evaluation method is expressed as follows: ； Where, is a feature quality indicator, is the neural network feature extraction function, is the similarity calculation function, Represents preliminary features, Indicates the preset label, The number of neural networks that feed the feature extraction module for the current batch; Step S26: According to the feature quality index, the parameters of the firefly algorithm are adjusted to more finely adjust the parameters of the neural network. The adjustment method is expressed as follows: ； Where, is the adjusted attractiveness coefficient; is the attractive learning rate, represents the exponential function; Step S27: Repeat steps S21 to S26 until a preset stopping condition is met.

2. The air-conditioning user comfort evaluation method considering meteorological influences according to claim 1, characterized in that: Loss function in step S22 Expressed as: ； Where, The number of neural networks that feed the feature extraction module for the current batch, Indicates the category with the highest probability obtained by calculating the preset Softmax classification function based on the neural network prediction output of the feature extraction module; represents the true label of the comfort data; is the counterfactual regularization term; is the regularization parameter of the neural network of the feature extraction module; is the weight matrix The Frobenius norm of .

3. The air-conditioning user comfort evaluation method considering meteorological influences according to claim 1, characterized in that: In step S24, the optimal structure is determined by cross-validation. Fold cross validation, expressed as: ； Where, Represents the k-fold cross validation under a given weight matrix Wp and the k-fold comfort evaluation dataset The performance evaluation value under Indicates the The number of comfort evaluation data points in the comfort evaluation dataset; Represents the loss function for neural network architecture search; Is the neural network in a given weight matrix Lower pair input The output of is calculated by the preset Softmax classification function to obtain the category with the highest probability. Represents the output of the neural network.

4. The method for evaluating air-conditioning user comfort considering meteorological influences according to claim 1, wherein: Step S23 specifically includes: constructing a differentiable causal graph network, defining structured causal equations between multi-source data of source collected information, and explicitly modeling causal paths through a gating mechanism; specifically: Get and preset weather data , the air conditioning setting parameters are , the physiological response data is ; in, is the meteorological data dimension; Set the dimensions for the air conditioner, Dimensions representing physiological response data; represents the set of real numbers; The causal propagation process of the differentiable causal graph network is modeled as: ； ； ； ； Where, The weight matrix of the causal path from meteorological data to air conditioning, The bias term for the causal path from meteorological data to air conditioning, The weight matrix for the causal paths from conditioning to physiological response data, A bias term that sets the causal path to the physiological response data for conditioning, is the gating weight matrix, is the gate dimension, is the gate bias term, is the ReLU activation function, is the Sigmoid function, is the vector concatenation operation, is element-wise multiplication, is a dynamic gating vector used to control the weight of the indirect impact of meteorological data on physiological response data; represents the causal inference fusion gate; The pseudo-correlation suppression mechanism constrains the feature space through causal intervention loss, eliminates the influence of confounding factors, and defines the counterfactual regularization term as: ； Where, is the comfort prediction value of the neural network, For meteorological data The Jacobian matrix of the preliminary features, Set up the air conditioner The Jacobian matrix of the preliminary features, is the Frobenius norm, the square root of the sum of the squares of the matrix elements, is the regularization penalty coefficient, represents the set of positive real numbers; By constraining the direct gradient contribution of meteorological data and air conditioning settings to the initial features of the output, the differentiable causal graph network is made to be only based on physiological response data. Predict the causal path to block the non-causal association path; Multi-source feature fusion is performed by performing a residual connection between the output of the differentiable causal graph network and the original features, preserving the complementary information of explicit causal features and implicit correlation features. Through the dynamic fusion coefficient, the contribution of explicit causal reasoning and data-driven implicit features is balanced to avoid oversimplification of real scenarios by causal modeling. It is expressed as: ； Where, is the original input feature, It is a multi-layer perceptron used to extract implicit features; is the learnable dynamic fusion coefficient, and the update method is , is the learning rate, is the feature quality indicator, is the final output of the differentiable causal graph network, that is, the preliminary feature.

5. The air-conditioning user comfort evaluation method considering meteorological influences according to claim 1, characterized in that: Step S3 is specifically as follows: The preliminary features are input into the prediction module, which is a fully connected layer. The prediction probability of each comfort level is output through the Softmax classification function. The category with the highest probability is used as the final comfort evaluation result to reflect the user's comfort status in the current air-conditioning environment.

Images